⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.
clm-0109
measured-herehigh ●●●
citable URL: https://halobench.com/records/clm-0109/ — this address never moves; the anchor /records/#clm-0109 keeps resolving
Null-update diagnostic: the v0.5.x base-platform Stage-A sentinel empty-content defect is RESOLVED on the reviewed v0.6.10 portable Vulkan/RADV runtime (2586f6edd). The plain/control arm that failed 6/15 Stage-A sentinel on v0.5.x (run-0541/0542/0543) PASSES on v0.6.10 (0 empty-assistant, 0 empty-tool-call, 0 single-token-EOS, code-valid 3/3, tool-correct 3/3) plus the 4/4 house guard, so the plain-vs-DFlash2-vs-stock-MTP comparison becomes legal. Both spec arms (stock-MTP n_max=3, DFlash2 Q8_0 n_max=3) also pass Stage-A sentinel and the 4/4 house guard with log-backed draft activation, but the G2 varied-prompt invariance check is NEGATIVE for both — neither is byte-invariant to the plain control (2/5 prompt classes match; DFlash2 first-diff on code, stock-MTP first-diff on toolish) — so both spec arms are recorded as NOT PROMOTED.
HO-004-DIAG / clm-0098 successor on v0.6.10. Scope boundary: Stage-A sentinel and house-guard pass only. NOT "DFlash2 works", NOT "DFlash2 beats MTP", NOT a quality-equivalence or throughput-improvement claim, NOT a production recommendation — the exactness gate (G2) is negative for both spec arms, so nothing beyond the empty-content defect resolution is admitted. Deviations: ran Vulkan/RADV per hborchestrator (card ROCm line was boilerplate); DFlash2 n_max=3 because the card forbids n_max>=4; G2 first attempt crashed on a runner slicing bug and re-ran G2-only. Energy: HA wall-meter join is deferred to the energy-join owner (runner lane had no HA MCP) with per-arm windows recorded; join before the ~10-day HA retention closes. No Wh/correct anywhere. This is a sentinel/gate diagnostic, not a capability or throughput benchmark.